| Date | Commit message (Collapse) |
|---|
|
We know that all versions of MRI have a small RFile structure
that is allocated in the same object slots as other Ruby types
and also zeroed on allocation.
This optimization enables us to fall back to using ivars in
case MRI changes or if we're used on other Rubies.
|
|
Hopefully it works for people who use TCP_NOPUSH...
|
|
Duh...
|
|
This is probably a better name for it, libautocork is a nice
name even though we won't use it directly.
|
|
It only supports TCP_CORK under Linux right now.
We use a very basic strategy to use TCP_CORK semantics optimally
in most TCP servers: On corked sockets, we will uncork on recv()
if there was a previous send(). Otherwise we do not fiddle
with TCP_CORK at all.
Under Linux, we can rely on TCP_CORK being inherited in an
accept()-ed client socket so we can avoid syscalls for each
accept()-ed client if we already know the accept() socket corks.
This module does NOTHING for client TCP sockets, we only deal
with accept()-ed sockets right now.
|
|
More common code that's still GNU make is better for my
sanity. Also, bogomips.org went on a URL diet recently.
|
|
There could be some platforms that dislike it...
|
|
Oops
|
|
We now avoid errno side-effects in kgio_wait_*able methods.
This affects Rubinius, but may affect other Ruby platforms
(particularly those that use stdio) as well.
|
|
Retrieving the file descriptor may have side-effects on
certain Ruby implementations (e.g. Rubinius), so ensure
our errno is preserved before calling rb_io_wait_*able().
|
|
It's more useful that way.
|
|
kgio_accept and kgio_tryaccept now take an optional argument
to override the default Kgio::Socket class that is returned.
These methods also fall back to using regular accept() if
kgio was built on a system with accept4() and later run on
a system without accept4().
|
|
Less noise means we'll notice real bugs sooner.
|
|
kgio may occasionally be built on a system with accept4()
and then deployed on one without it. Handle this case
gracefully since it unfortunately happens on production systems.
|
|
This is preferred as we no longer have to rely on a global
constant.
|
|
Oops, we were never Ruby licensed.
|
|
wrongdoc factors out a bunch of common code from this
project into its own and removes JavaScript from RDoc
to boot.
|
|
Noticed-by: Iñaki Baz Castillo
|
|
(no code changes from 2.0.0pre1)
This release should make Kgio easier and more consistent
to use across a variety of libraries/applications.
The global Kgio.wait_*able(=) accessor methods are gone in favor
of having default kgio_wait_readable and kgio_wait_writable
methods added to all Kgio-using classes. Sub-classes may (and
are encouraged to) redefine these if needed.
Eric Wong (7):
expand Kgio::*#kgio_read! documentation
prefer symbolic names for waiting read/writability
EOFError message matches Ruby's
README: Gemcutter => RubyGems.org
update documentation with mailing list info
add default kgio_wait_*able methods
switch entirely to kgio_wait_*able methods
|
|
Since we do prerelease nowadays before real ones.
|
|
This project is useful enough for others and to stand alone
without needing to be associated with Unicorn.
|
|
This release should make Kgio easier and more consistent
to use across a variety of libraries/applications.
The global Kgio.wait_*able(=) accessor methods are gone in favor
of having default kgio_wait_readable and kgio_wait_writable
methods added to all Kgio-using classes. Sub-classes may (and
are encouraged to) redefine these if needed.
Eric Wong (7):
expand Kgio::*#kgio_read! documentation
prefer symbolic names for waiting read/writability
EOFError message matches Ruby's
README: Gemcutter => RubyGems.org
update documentation with mailing list info
add default kgio_wait_*able methods
switch entirely to kgio_wait_*able methods
|
|
This removes the global Kgio.wait_*able accesors and requires
each class to define (or fall back to) the Kgio::DefaultWaiters
methods.
|
|
It makes it easier for people to use certain overrides without
killing other methods. This is the first step in fixing
problems people were having with dalli 0.11.1+ while running
Unicorn.
|
|
We're a real project, apparently, so it can have its
own mailing list.
|
|
That's the new name for it and it's official
|
|
This makes messages appear less different than Ruby
when using kgio_read!
Requested-by: Mike Perham
|
|
There's no point in using constants that point to symbols
instead of just the symbols themselves.
|
|
If the author can forget why it was written, so can
the rest of the world.
|
|
kgio_read and kgio_tryread will now return an empty string when
a length of zero is specified instead of nil (which would signal
an EOF). This emulates the behavior of IO#read, IO#readpartial,
IO#sysread, IO#read_nonblock in core Ruby for consistency.
|
|
This matches behavior of all the core Ruby methods.
|
|
* make Kgio::WaitWritable and Kgio::WaitReadable symbols
* trywrite: fix stupid off-by-one error causing corrupt writes
on retries
|
|
case/when and === didn't actually work as I expected
them to.
|
|
Oops!
|
|
This makes them easier to compare with === when used
in case/when statements in Ruby
|
|
This fixes our accept4() wrapper which did not work as expected
on some *BSD-based systems due to fcntl(fd, F_GETFL) returning
false information. Linux 2.6+ users are unnaffected, including
those without accept4().
Also some RDoc fixes.
|
|
Oops, completely broken by the splitting of the code.
|
|
Apparently fcntl(fd, F_GETFL) can return falsely return the
O_NONBLOCK flag without actually having it set in the kernel.
This is totally broken on the part of the OS.
|
|
We build more than one file nowadays.
|
|
oops...
|
|
The C extension is now split into several files for
ease-of-maintenance.
Slightly more common, client-triggerable exceptions (EOFError,
Errno::EPIPE, Errno::ECONNRESET) are now less expensive as they
are generated without backtraces.
|
|
Malicious clients may disconnect during big writes to cause
EPIPE and ECONNRESET exceptions. Generating backtraces can be
expensive with Ruby, so mitigate the DoS vector by lowering the
cost of generating an exception.
|
|
Except EOFError is gently raised to not include a huge
backtrace. Large backtraces can be a performance problem on
busy servers that malicious clients may exploit to deny service.
|
|
We've been spoiled by GNU tar.
|
|
Making the code easier to read and navigate. This also
frees us from having to use the stupid A4_ prefix for
accept4(2) flags since it conflicts with the socket(2)
ones.
|
|
Partial writes can be retried until completely denied with
EAGAIN. Often times, it is beneficial to retry immediately
after a partial write because the kernel may allocate more
buffers or the reader can drain the buffers.
This helps the caller avoid crossing the Ruby <-> C boundary
more than necessary.
|
|
* an alternate class now be returned by accept/tryaccept
by setting "Kgio.accept_class ="
|
|
These can be useful for avoiding wrapper objects and
also allows users to more easily try different things
without stepping on others' toe^H^H^Hclasses.
|
|
* add compatibility for ancient Rubies (1.8.6)
* linux: fix accept4() support for newer Linux
|
|
Oops :x Tested on Debian sid.
|